深度学习论文阅读笔记(1)-VGG

VERY DEEP CONVOLUTIONAL NETWORKS FOR LARGE-SCALE IMAGE RECOGNITION

2014 1st, and 2nd in the localisation and classification ImageNet Challenge

1. small receptive field \(3\times 3\) is better than larger filter, and it is kind of regualarization. Smaller stride is used too, and this way can capture more information.
2. \(1\times 1\) convolution is utilised, and it pay all attention to one pixel without caring its neightborhood, and it can change channel size easily
3. smaller maxpooling is used,and size is 2*2, stride is 2
4. In testing period, change the final FC layers to conv layers
   
5. layers are deeper and wider